Fix: Interesting how well it competes against Qwen3.6 35B and was Qwen3.6 retrained or Qwen3.5

.gitignore

@@ -0,0 +1,4 @@
+__pycache__/
+*.py[cod]
+*$py.class
+.pytest_cache/

README.md

@@ -15,7 +15,7 @@
   <img src="assets/group_capybaras_flat.png" alt="Qwen-AgentWorld" width="1000px"/>
 </div>
 
-Welcome to the GitHub repository of Qwen-AgentWorld. Here, you can find official information about Qwen-AgentWorld, post your questions ([Issues](https://github.com/QwenLM/Qwen-AgentWorld/issues)), and share your ideas with the community ([Discussions](https://github.com/QwenLM/Qwen-AgentWorld/discussions)).
+Welcome to the GitHub repository of Qwen-AgentWorld. Here, you can find official information about Qwen-AgentWorld, post your questions ([Issues](https://github.com/QwenLM/Qwen-AgentWorld/issues)), and share your ideas with the community ([Discussions](https://github.com/QwenLM/Qwen-AgentWorld/discussions)). See the [FAQ](docs/FAQ.md) for common questions about benchmark comparisons and evaluation prompts.
 
 
 ## News

docs/FAQ.md

@@ -0,0 +1,17 @@
+# FAQ
+
+## Model comparison and training lineage
+
+Qwen-AgentWorld benchmark results can be compared with other published model results under the evaluation setup described in the README, blog, technical report, or model card. This repository contains the AgentWorldBench evaluation tooling and prompt/configuration code. It is not the authoritative source for unpublished third-party model lineage or retraining history.
+
+For model-background questions, please rely on the model card, release notes, or maintainer announcements for the specific checkpoint being evaluated. If a benchmark table compares against another model family, treat it as an evaluation result under the stated benchmark setup rather than a statement about that model's training recipe unless the release explicitly says so.
+
+## Official prompt for SimRL/OpenClaw-style evaluation
+
+The evaluation script does not hard-code one universal prompt for SimRL or OpenClaw-style runs. It builds chat messages from each benchmark sample:
+
+- `system_str`, when present, is sent as the system message.
+- `current_prompt`, when present, is sent as the current user message.
+- If `current_prompt` is absent, the script falls back to the sample's `prompt` list and `turn_idx`.
+
+Use the prompt fields shipped with the benchmark/evaluation data as the canonical prompt source for reproducing AgentWorldBench runs. Domain-specific response markers and prompt file paths are configured in `eval/lwm_eval_utils/task_configs.py`.

tests/test_faq.py

@@ -0,0 +1,20 @@
+import unittest
+from pathlib import Path
+
+
+class FaqDocumentationTest(unittest.TestCase):
+    def test_faq_documents_model_lineage_scope_and_prompt_source(self):
+        faq = Path(__file__).resolve().parents[1] / "docs" / "FAQ.md"
+        text = faq.read_text(encoding="utf-8")
+
+        self.assertIn("benchmark results", text)
+        self.assertIn("not the authoritative source", text)
+        self.assertIn("third-party model lineage", text)
+        self.assertIn("SimRL or OpenClaw-style runs", text)
+        self.assertIn("system_str", text)
+        self.assertIn("current_prompt", text)
+        self.assertIn("prompt fields shipped with the benchmark/evaluation data", text)
+
+
+if __name__ == "__main__":
+    unittest.main()